1 #define JEMALLOC_PROF_C_
2 #include "jemalloc/internal/jemalloc_internal.h"
3 /******************************************************************************/
4
5 #ifdef JEMALLOC_PROF_LIBUNWIND
6 #define UNW_LOCAL_ONLY
7 #include <libunwind.h>
8 #endif
9
10 #ifdef JEMALLOC_PROF_LIBGCC
11 #include <unwind.h>
12 #endif
13
14 /******************************************************************************/
15 /* Data. */
16
17 bool opt_prof = false;
18 bool opt_prof_active = true;
19 bool opt_prof_thread_active_init = true;
20 size_t opt_lg_prof_sample = LG_PROF_SAMPLE_DEFAULT;
21 ssize_t opt_lg_prof_interval = LG_PROF_INTERVAL_DEFAULT;
22 bool opt_prof_gdump = false;
23 bool opt_prof_final = false;
24 bool opt_prof_leak = false;
25 bool opt_prof_accum = false;
26 char opt_prof_prefix[
27 /* Minimize memory bloat for non-prof builds. */
28 #ifdef JEMALLOC_PROF
29 PATH_MAX +
30 #endif
31 1];
32
33 /*
34 * Initialized as opt_prof_active, and accessed via
35 * prof_active_[gs]et{_unlocked,}().
36 */
37 bool prof_active;
38 static malloc_mutex_t prof_active_mtx;
39
40 /*
41 * Initialized as opt_prof_thread_active_init, and accessed via
42 * prof_thread_active_init_[gs]et().
43 */
44 static bool prof_thread_active_init;
45 static malloc_mutex_t prof_thread_active_init_mtx;
46
47 /*
48 * Initialized as opt_prof_gdump, and accessed via
49 * prof_gdump_[gs]et{_unlocked,}().
50 */
51 bool prof_gdump_val;
52 static malloc_mutex_t prof_gdump_mtx;
53
54 uint64_t prof_interval = 0;
55
56 size_t lg_prof_sample;
57
58 /*
59 * Table of mutexes that are shared among gctx's. These are leaf locks, so
60 * there is no problem with using them for more than one gctx at the same time.
61 * The primary motivation for this sharing though is that gctx's are ephemeral,
62 * and destroying mutexes causes complications for systems that allocate when
63 * creating/destroying mutexes.
64 */
65 static malloc_mutex_t *gctx_locks;
66 static unsigned cum_gctxs; /* Atomic counter. */
67
68 /*
69 * Table of mutexes that are shared among tdata's. No operations require
70 * holding multiple tdata locks, so there is no problem with using them for more
71 * than one tdata at the same time, even though a gctx lock may be acquired
72 * while holding a tdata lock.
73 */
74 static malloc_mutex_t *tdata_locks;
75
76 /*
77 * Global hash of (prof_bt_t *)-->(prof_gctx_t *). This is the master data
78 * structure that knows about all backtraces currently captured.
79 */
80 static ckh_t bt2gctx;
81 static malloc_mutex_t bt2gctx_mtx;
82
83 /*
84 * Tree of all extant prof_tdata_t structures, regardless of state,
85 * {attached,detached,expired}.
86 */
87 static prof_tdata_tree_t tdatas;
88 static malloc_mutex_t tdatas_mtx;
89
90 static uint64_t next_thr_uid;
91 static malloc_mutex_t next_thr_uid_mtx;
92
93 static malloc_mutex_t prof_dump_seq_mtx;
94 static uint64_t prof_dump_seq;
95 static uint64_t prof_dump_iseq;
96 static uint64_t prof_dump_mseq;
97 static uint64_t prof_dump_useq;
98
99 /*
100 * This buffer is rather large for stack allocation, so use a single buffer for
101 * all profile dumps.
102 */
103 static malloc_mutex_t prof_dump_mtx;
104 static char prof_dump_buf[
105 /* Minimize memory bloat for non-prof builds. */
106 #ifdef JEMALLOC_PROF
107 PROF_DUMP_BUFSIZE
108 #else
109 1
110 #endif
111 ];
112 static size_t prof_dump_buf_end;
113 static int prof_dump_fd;
114
115 /* Do not dump any profiles until bootstrapping is complete. */
116 static bool prof_booted = false;
117
118 /******************************************************************************/
119 /*
120 * Function prototypes for static functions that are referenced prior to
121 * definition.
122 */
123
124 static bool prof_tctx_should_destroy(prof_tctx_t *tctx);
125 static void prof_tctx_destroy(tsd_t *tsd, prof_tctx_t *tctx);
126 static bool prof_tdata_should_destroy(prof_tdata_t *tdata,
127 bool even_if_attached);
128 static void prof_tdata_destroy(tsd_t *tsd, prof_tdata_t *tdata,
129 bool even_if_attached);
130 static char *prof_thread_name_alloc(tsd_t *tsd, const char *thread_name);
131
132 /******************************************************************************/
133 /* Red-black trees. */
134
135 JEMALLOC_INLINE_C int
prof_tctx_comp(const prof_tctx_t * a,const prof_tctx_t * b)136 prof_tctx_comp(const prof_tctx_t *a, const prof_tctx_t *b)
137 {
138 uint64_t a_thr_uid = a->thr_uid;
139 uint64_t b_thr_uid = b->thr_uid;
140 int ret = (a_thr_uid > b_thr_uid) - (a_thr_uid < b_thr_uid);
141 if (ret == 0) {
142 uint64_t a_thr_discrim = a->thr_discrim;
143 uint64_t b_thr_discrim = b->thr_discrim;
144 ret = (a_thr_discrim > b_thr_discrim) - (a_thr_discrim <
145 b_thr_discrim);
146 if (ret == 0) {
147 uint64_t a_tctx_uid = a->tctx_uid;
148 uint64_t b_tctx_uid = b->tctx_uid;
149 ret = (a_tctx_uid > b_tctx_uid) - (a_tctx_uid <
150 b_tctx_uid);
151 }
152 }
153 return (ret);
154 }
155
rb_gen(static UNUSED,tctx_tree_,prof_tctx_tree_t,prof_tctx_t,tctx_link,prof_tctx_comp)156 rb_gen(static UNUSED, tctx_tree_, prof_tctx_tree_t, prof_tctx_t,
157 tctx_link, prof_tctx_comp)
158
159 JEMALLOC_INLINE_C int
160 prof_gctx_comp(const prof_gctx_t *a, const prof_gctx_t *b)
161 {
162 unsigned a_len = a->bt.len;
163 unsigned b_len = b->bt.len;
164 unsigned comp_len = (a_len < b_len) ? a_len : b_len;
165 int ret = memcmp(a->bt.vec, b->bt.vec, comp_len * sizeof(void *));
166 if (ret == 0)
167 ret = (a_len > b_len) - (a_len < b_len);
168 return (ret);
169 }
170
rb_gen(static UNUSED,gctx_tree_,prof_gctx_tree_t,prof_gctx_t,dump_link,prof_gctx_comp)171 rb_gen(static UNUSED, gctx_tree_, prof_gctx_tree_t, prof_gctx_t, dump_link,
172 prof_gctx_comp)
173
174 JEMALLOC_INLINE_C int
175 prof_tdata_comp(const prof_tdata_t *a, const prof_tdata_t *b)
176 {
177 int ret;
178 uint64_t a_uid = a->thr_uid;
179 uint64_t b_uid = b->thr_uid;
180
181 ret = ((a_uid > b_uid) - (a_uid < b_uid));
182 if (ret == 0) {
183 uint64_t a_discrim = a->thr_discrim;
184 uint64_t b_discrim = b->thr_discrim;
185
186 ret = ((a_discrim > b_discrim) - (a_discrim < b_discrim));
187 }
188 return (ret);
189 }
190
rb_gen(static UNUSED,tdata_tree_,prof_tdata_tree_t,prof_tdata_t,tdata_link,prof_tdata_comp)191 rb_gen(static UNUSED, tdata_tree_, prof_tdata_tree_t, prof_tdata_t, tdata_link,
192 prof_tdata_comp)
193
194 /******************************************************************************/
195
196 void
197 prof_alloc_rollback(tsd_t *tsd, prof_tctx_t *tctx, bool updated)
198 {
199 prof_tdata_t *tdata;
200
201 cassert(config_prof);
202
203 if (updated) {
204 /*
205 * Compute a new sample threshold. This isn't very important in
206 * practice, because this function is rarely executed, so the
207 * potential for sample bias is minimal except in contrived
208 * programs.
209 */
210 tdata = prof_tdata_get(tsd, true);
211 if (tdata != NULL)
212 prof_sample_threshold_update(tdata);
213 }
214
215 if ((uintptr_t)tctx > (uintptr_t)1U) {
216 malloc_mutex_lock(tctx->tdata->lock);
217 tctx->prepared = false;
218 if (prof_tctx_should_destroy(tctx))
219 prof_tctx_destroy(tsd, tctx);
220 else
221 malloc_mutex_unlock(tctx->tdata->lock);
222 }
223 }
224
225 void
prof_malloc_sample_object(const void * ptr,size_t usize,prof_tctx_t * tctx)226 prof_malloc_sample_object(const void *ptr, size_t usize, prof_tctx_t *tctx)
227 {
228
229 prof_tctx_set(ptr, usize, tctx);
230
231 malloc_mutex_lock(tctx->tdata->lock);
232 tctx->cnts.curobjs++;
233 tctx->cnts.curbytes += usize;
234 if (opt_prof_accum) {
235 tctx->cnts.accumobjs++;
236 tctx->cnts.accumbytes += usize;
237 }
238 tctx->prepared = false;
239 malloc_mutex_unlock(tctx->tdata->lock);
240 }
241
242 void
prof_free_sampled_object(tsd_t * tsd,size_t usize,prof_tctx_t * tctx)243 prof_free_sampled_object(tsd_t *tsd, size_t usize, prof_tctx_t *tctx)
244 {
245
246 malloc_mutex_lock(tctx->tdata->lock);
247 assert(tctx->cnts.curobjs > 0);
248 assert(tctx->cnts.curbytes >= usize);
249 tctx->cnts.curobjs--;
250 tctx->cnts.curbytes -= usize;
251
252 if (prof_tctx_should_destroy(tctx))
253 prof_tctx_destroy(tsd, tctx);
254 else
255 malloc_mutex_unlock(tctx->tdata->lock);
256 }
257
258 void
bt_init(prof_bt_t * bt,void ** vec)259 bt_init(prof_bt_t *bt, void **vec)
260 {
261
262 cassert(config_prof);
263
264 bt->vec = vec;
265 bt->len = 0;
266 }
267
268 JEMALLOC_INLINE_C void
prof_enter(tsd_t * tsd,prof_tdata_t * tdata)269 prof_enter(tsd_t *tsd, prof_tdata_t *tdata)
270 {
271
272 cassert(config_prof);
273 assert(tdata == prof_tdata_get(tsd, false));
274
275 if (tdata != NULL) {
276 assert(!tdata->enq);
277 tdata->enq = true;
278 }
279
280 malloc_mutex_lock(&bt2gctx_mtx);
281 }
282
283 JEMALLOC_INLINE_C void
prof_leave(tsd_t * tsd,prof_tdata_t * tdata)284 prof_leave(tsd_t *tsd, prof_tdata_t *tdata)
285 {
286
287 cassert(config_prof);
288 assert(tdata == prof_tdata_get(tsd, false));
289
290 malloc_mutex_unlock(&bt2gctx_mtx);
291
292 if (tdata != NULL) {
293 bool idump, gdump;
294
295 assert(tdata->enq);
296 tdata->enq = false;
297 idump = tdata->enq_idump;
298 tdata->enq_idump = false;
299 gdump = tdata->enq_gdump;
300 tdata->enq_gdump = false;
301
302 if (idump)
303 prof_idump();
304 if (gdump)
305 prof_gdump();
306 }
307 }
308
309 #ifdef JEMALLOC_PROF_LIBUNWIND
310 void
prof_backtrace(prof_bt_t * bt)311 prof_backtrace(prof_bt_t *bt)
312 {
313 int nframes;
314
315 cassert(config_prof);
316 assert(bt->len == 0);
317 assert(bt->vec != NULL);
318
319 nframes = unw_backtrace(bt->vec, PROF_BT_MAX);
320 if (nframes <= 0)
321 return;
322 bt->len = nframes;
323 }
324 #elif (defined(JEMALLOC_PROF_LIBGCC))
325 static _Unwind_Reason_Code
prof_unwind_init_callback(struct _Unwind_Context * context,void * arg)326 prof_unwind_init_callback(struct _Unwind_Context *context, void *arg)
327 {
328
329 cassert(config_prof);
330
331 return (_URC_NO_REASON);
332 }
333
334 static _Unwind_Reason_Code
prof_unwind_callback(struct _Unwind_Context * context,void * arg)335 prof_unwind_callback(struct _Unwind_Context *context, void *arg)
336 {
337 prof_unwind_data_t *data = (prof_unwind_data_t *)arg;
338 void *ip;
339
340 cassert(config_prof);
341
342 ip = (void *)_Unwind_GetIP(context);
343 if (ip == NULL)
344 return (_URC_END_OF_STACK);
345 data->bt->vec[data->bt->len] = ip;
346 data->bt->len++;
347 if (data->bt->len == data->max)
348 return (_URC_END_OF_STACK);
349
350 return (_URC_NO_REASON);
351 }
352
353 void
prof_backtrace(prof_bt_t * bt)354 prof_backtrace(prof_bt_t *bt)
355 {
356 prof_unwind_data_t data = {bt, PROF_BT_MAX};
357
358 cassert(config_prof);
359
360 _Unwind_Backtrace(prof_unwind_callback, &data);
361 }
362 #elif (defined(JEMALLOC_PROF_GCC))
363 void
prof_backtrace(prof_bt_t * bt)364 prof_backtrace(prof_bt_t *bt)
365 {
366 #define BT_FRAME(i) \
367 if ((i) < PROF_BT_MAX) { \
368 void *p; \
369 if (__builtin_frame_address(i) == 0) \
370 return; \
371 p = __builtin_return_address(i); \
372 if (p == NULL) \
373 return; \
374 bt->vec[(i)] = p; \
375 bt->len = (i) + 1; \
376 } else \
377 return;
378
379 cassert(config_prof);
380
381 BT_FRAME(0)
382 BT_FRAME(1)
383 BT_FRAME(2)
384 BT_FRAME(3)
385 BT_FRAME(4)
386 BT_FRAME(5)
387 BT_FRAME(6)
388 BT_FRAME(7)
389 BT_FRAME(8)
390 BT_FRAME(9)
391
392 BT_FRAME(10)
393 BT_FRAME(11)
394 BT_FRAME(12)
395 BT_FRAME(13)
396 BT_FRAME(14)
397 BT_FRAME(15)
398 BT_FRAME(16)
399 BT_FRAME(17)
400 BT_FRAME(18)
401 BT_FRAME(19)
402
403 BT_FRAME(20)
404 BT_FRAME(21)
405 BT_FRAME(22)
406 BT_FRAME(23)
407 BT_FRAME(24)
408 BT_FRAME(25)
409 BT_FRAME(26)
410 BT_FRAME(27)
411 BT_FRAME(28)
412 BT_FRAME(29)
413
414 BT_FRAME(30)
415 BT_FRAME(31)
416 BT_FRAME(32)
417 BT_FRAME(33)
418 BT_FRAME(34)
419 BT_FRAME(35)
420 BT_FRAME(36)
421 BT_FRAME(37)
422 BT_FRAME(38)
423 BT_FRAME(39)
424
425 BT_FRAME(40)
426 BT_FRAME(41)
427 BT_FRAME(42)
428 BT_FRAME(43)
429 BT_FRAME(44)
430 BT_FRAME(45)
431 BT_FRAME(46)
432 BT_FRAME(47)
433 BT_FRAME(48)
434 BT_FRAME(49)
435
436 BT_FRAME(50)
437 BT_FRAME(51)
438 BT_FRAME(52)
439 BT_FRAME(53)
440 BT_FRAME(54)
441 BT_FRAME(55)
442 BT_FRAME(56)
443 BT_FRAME(57)
444 BT_FRAME(58)
445 BT_FRAME(59)
446
447 BT_FRAME(60)
448 BT_FRAME(61)
449 BT_FRAME(62)
450 BT_FRAME(63)
451 BT_FRAME(64)
452 BT_FRAME(65)
453 BT_FRAME(66)
454 BT_FRAME(67)
455 BT_FRAME(68)
456 BT_FRAME(69)
457
458 BT_FRAME(70)
459 BT_FRAME(71)
460 BT_FRAME(72)
461 BT_FRAME(73)
462 BT_FRAME(74)
463 BT_FRAME(75)
464 BT_FRAME(76)
465 BT_FRAME(77)
466 BT_FRAME(78)
467 BT_FRAME(79)
468
469 BT_FRAME(80)
470 BT_FRAME(81)
471 BT_FRAME(82)
472 BT_FRAME(83)
473 BT_FRAME(84)
474 BT_FRAME(85)
475 BT_FRAME(86)
476 BT_FRAME(87)
477 BT_FRAME(88)
478 BT_FRAME(89)
479
480 BT_FRAME(90)
481 BT_FRAME(91)
482 BT_FRAME(92)
483 BT_FRAME(93)
484 BT_FRAME(94)
485 BT_FRAME(95)
486 BT_FRAME(96)
487 BT_FRAME(97)
488 BT_FRAME(98)
489 BT_FRAME(99)
490
491 BT_FRAME(100)
492 BT_FRAME(101)
493 BT_FRAME(102)
494 BT_FRAME(103)
495 BT_FRAME(104)
496 BT_FRAME(105)
497 BT_FRAME(106)
498 BT_FRAME(107)
499 BT_FRAME(108)
500 BT_FRAME(109)
501
502 BT_FRAME(110)
503 BT_FRAME(111)
504 BT_FRAME(112)
505 BT_FRAME(113)
506 BT_FRAME(114)
507 BT_FRAME(115)
508 BT_FRAME(116)
509 BT_FRAME(117)
510 BT_FRAME(118)
511 BT_FRAME(119)
512
513 BT_FRAME(120)
514 BT_FRAME(121)
515 BT_FRAME(122)
516 BT_FRAME(123)
517 BT_FRAME(124)
518 BT_FRAME(125)
519 BT_FRAME(126)
520 BT_FRAME(127)
521 #undef BT_FRAME
522 }
523 #else
524 void
prof_backtrace(prof_bt_t * bt)525 prof_backtrace(prof_bt_t *bt)
526 {
527
528 cassert(config_prof);
529 not_reached();
530 }
531 #endif
532
533 static malloc_mutex_t *
prof_gctx_mutex_choose(void)534 prof_gctx_mutex_choose(void)
535 {
536 unsigned ngctxs = atomic_add_u(&cum_gctxs, 1);
537
538 return (&gctx_locks[(ngctxs - 1) % PROF_NCTX_LOCKS]);
539 }
540
541 static malloc_mutex_t *
prof_tdata_mutex_choose(uint64_t thr_uid)542 prof_tdata_mutex_choose(uint64_t thr_uid)
543 {
544
545 return (&tdata_locks[thr_uid % PROF_NTDATA_LOCKS]);
546 }
547
548 static prof_gctx_t *
prof_gctx_create(tsd_t * tsd,prof_bt_t * bt)549 prof_gctx_create(tsd_t *tsd, prof_bt_t *bt)
550 {
551 /*
552 * Create a single allocation that has space for vec of length bt->len.
553 */
554 size_t size = offsetof(prof_gctx_t, vec) + (bt->len * sizeof(void *));
555 prof_gctx_t *gctx = (prof_gctx_t *)iallocztm(tsd, size,
556 size2index(size), false, tcache_get(tsd, true), true, NULL, true);
557 if (gctx == NULL)
558 return (NULL);
559 gctx->lock = prof_gctx_mutex_choose();
560 /*
561 * Set nlimbo to 1, in order to avoid a race condition with
562 * prof_tctx_destroy()/prof_gctx_try_destroy().
563 */
564 gctx->nlimbo = 1;
565 tctx_tree_new(&gctx->tctxs);
566 /* Duplicate bt. */
567 memcpy(gctx->vec, bt->vec, bt->len * sizeof(void *));
568 gctx->bt.vec = gctx->vec;
569 gctx->bt.len = bt->len;
570 return (gctx);
571 }
572
573 static void
prof_gctx_try_destroy(tsd_t * tsd,prof_tdata_t * tdata_self,prof_gctx_t * gctx,prof_tdata_t * tdata)574 prof_gctx_try_destroy(tsd_t *tsd, prof_tdata_t *tdata_self, prof_gctx_t *gctx,
575 prof_tdata_t *tdata)
576 {
577
578 cassert(config_prof);
579
580 /*
581 * Check that gctx is still unused by any thread cache before destroying
582 * it. prof_lookup() increments gctx->nlimbo in order to avoid a race
583 * condition with this function, as does prof_tctx_destroy() in order to
584 * avoid a race between the main body of prof_tctx_destroy() and entry
585 * into this function.
586 */
587 prof_enter(tsd, tdata_self);
588 malloc_mutex_lock(gctx->lock);
589 assert(gctx->nlimbo != 0);
590 if (tctx_tree_empty(&gctx->tctxs) && gctx->nlimbo == 1) {
591 /* Remove gctx from bt2gctx. */
592 if (ckh_remove(tsd, &bt2gctx, &gctx->bt, NULL, NULL))
593 not_reached();
594 prof_leave(tsd, tdata_self);
595 /* Destroy gctx. */
596 malloc_mutex_unlock(gctx->lock);
597 idalloctm(tsd, gctx, tcache_get(tsd, false), true, true);
598 } else {
599 /*
600 * Compensate for increment in prof_tctx_destroy() or
601 * prof_lookup().
602 */
603 gctx->nlimbo--;
604 malloc_mutex_unlock(gctx->lock);
605 prof_leave(tsd, tdata_self);
606 }
607 }
608
609 /* tctx->tdata->lock must be held. */
610 static bool
prof_tctx_should_destroy(prof_tctx_t * tctx)611 prof_tctx_should_destroy(prof_tctx_t *tctx)
612 {
613
614 if (opt_prof_accum)
615 return (false);
616 if (tctx->cnts.curobjs != 0)
617 return (false);
618 if (tctx->prepared)
619 return (false);
620 return (true);
621 }
622
623 static bool
prof_gctx_should_destroy(prof_gctx_t * gctx)624 prof_gctx_should_destroy(prof_gctx_t *gctx)
625 {
626
627 if (opt_prof_accum)
628 return (false);
629 if (!tctx_tree_empty(&gctx->tctxs))
630 return (false);
631 if (gctx->nlimbo != 0)
632 return (false);
633 return (true);
634 }
635
636 /* tctx->tdata->lock is held upon entry, and released before return. */
637 static void
prof_tctx_destroy(tsd_t * tsd,prof_tctx_t * tctx)638 prof_tctx_destroy(tsd_t *tsd, prof_tctx_t *tctx)
639 {
640 prof_tdata_t *tdata = tctx->tdata;
641 prof_gctx_t *gctx = tctx->gctx;
642 bool destroy_tdata, destroy_tctx, destroy_gctx;
643
644 assert(tctx->cnts.curobjs == 0);
645 assert(tctx->cnts.curbytes == 0);
646 assert(!opt_prof_accum);
647 assert(tctx->cnts.accumobjs == 0);
648 assert(tctx->cnts.accumbytes == 0);
649
650 ckh_remove(tsd, &tdata->bt2tctx, &gctx->bt, NULL, NULL);
651 destroy_tdata = prof_tdata_should_destroy(tdata, false);
652 malloc_mutex_unlock(tdata->lock);
653
654 malloc_mutex_lock(gctx->lock);
655 switch (tctx->state) {
656 case prof_tctx_state_nominal:
657 tctx_tree_remove(&gctx->tctxs, tctx);
658 destroy_tctx = true;
659 if (prof_gctx_should_destroy(gctx)) {
660 /*
661 * Increment gctx->nlimbo in order to keep another
662 * thread from winning the race to destroy gctx while
663 * this one has gctx->lock dropped. Without this, it
664 * would be possible for another thread to:
665 *
666 * 1) Sample an allocation associated with gctx.
667 * 2) Deallocate the sampled object.
668 * 3) Successfully prof_gctx_try_destroy(gctx).
669 *
670 * The result would be that gctx no longer exists by the
671 * time this thread accesses it in
672 * prof_gctx_try_destroy().
673 */
674 gctx->nlimbo++;
675 destroy_gctx = true;
676 } else
677 destroy_gctx = false;
678 break;
679 case prof_tctx_state_dumping:
680 /*
681 * A dumping thread needs tctx to remain valid until dumping
682 * has finished. Change state such that the dumping thread will
683 * complete destruction during a late dump iteration phase.
684 */
685 tctx->state = prof_tctx_state_purgatory;
686 destroy_tctx = false;
687 destroy_gctx = false;
688 break;
689 default:
690 not_reached();
691 destroy_tctx = false;
692 destroy_gctx = false;
693 }
694 malloc_mutex_unlock(gctx->lock);
695 if (destroy_gctx) {
696 prof_gctx_try_destroy(tsd, prof_tdata_get(tsd, false), gctx,
697 tdata);
698 }
699
700 if (destroy_tdata)
701 prof_tdata_destroy(tsd, tdata, false);
702
703 if (destroy_tctx)
704 idalloctm(tsd, tctx, tcache_get(tsd, false), true, true);
705 }
706
707 static bool
prof_lookup_global(tsd_t * tsd,prof_bt_t * bt,prof_tdata_t * tdata,void ** p_btkey,prof_gctx_t ** p_gctx,bool * p_new_gctx)708 prof_lookup_global(tsd_t *tsd, prof_bt_t *bt, prof_tdata_t *tdata,
709 void **p_btkey, prof_gctx_t **p_gctx, bool *p_new_gctx)
710 {
711 union {
712 prof_gctx_t *p;
713 void *v;
714 } gctx;
715 union {
716 prof_bt_t *p;
717 void *v;
718 } btkey;
719 bool new_gctx;
720
721 prof_enter(tsd, tdata);
722 if (ckh_search(&bt2gctx, bt, &btkey.v, &gctx.v)) {
723 /* bt has never been seen before. Insert it. */
724 gctx.p = prof_gctx_create(tsd, bt);
725 if (gctx.v == NULL) {
726 prof_leave(tsd, tdata);
727 return (true);
728 }
729 btkey.p = &gctx.p->bt;
730 if (ckh_insert(tsd, &bt2gctx, btkey.v, gctx.v)) {
731 /* OOM. */
732 prof_leave(tsd, tdata);
733 idalloctm(tsd, gctx.v, tcache_get(tsd, false), true,
734 true);
735 return (true);
736 }
737 new_gctx = true;
738 } else {
739 /*
740 * Increment nlimbo, in order to avoid a race condition with
741 * prof_tctx_destroy()/prof_gctx_try_destroy().
742 */
743 malloc_mutex_lock(gctx.p->lock);
744 gctx.p->nlimbo++;
745 malloc_mutex_unlock(gctx.p->lock);
746 new_gctx = false;
747 }
748 prof_leave(tsd, tdata);
749
750 *p_btkey = btkey.v;
751 *p_gctx = gctx.p;
752 *p_new_gctx = new_gctx;
753 return (false);
754 }
755
756 prof_tctx_t *
prof_lookup(tsd_t * tsd,prof_bt_t * bt)757 prof_lookup(tsd_t *tsd, prof_bt_t *bt)
758 {
759 union {
760 prof_tctx_t *p;
761 void *v;
762 } ret;
763 prof_tdata_t *tdata;
764 bool not_found;
765
766 cassert(config_prof);
767
768 tdata = prof_tdata_get(tsd, false);
769 if (tdata == NULL)
770 return (NULL);
771
772 malloc_mutex_lock(tdata->lock);
773 not_found = ckh_search(&tdata->bt2tctx, bt, NULL, &ret.v);
774 if (!not_found) /* Note double negative! */
775 ret.p->prepared = true;
776 malloc_mutex_unlock(tdata->lock);
777 if (not_found) {
778 tcache_t *tcache;
779 void *btkey;
780 prof_gctx_t *gctx;
781 bool new_gctx, error;
782
783 /*
784 * This thread's cache lacks bt. Look for it in the global
785 * cache.
786 */
787 if (prof_lookup_global(tsd, bt, tdata, &btkey, &gctx,
788 &new_gctx))
789 return (NULL);
790
791 /* Link a prof_tctx_t into gctx for this thread. */
792 tcache = tcache_get(tsd, true);
793 ret.v = iallocztm(tsd, sizeof(prof_tctx_t),
794 size2index(sizeof(prof_tctx_t)), false, tcache, true, NULL,
795 true);
796 if (ret.p == NULL) {
797 if (new_gctx)
798 prof_gctx_try_destroy(tsd, tdata, gctx, tdata);
799 return (NULL);
800 }
801 ret.p->tdata = tdata;
802 ret.p->thr_uid = tdata->thr_uid;
803 ret.p->thr_discrim = tdata->thr_discrim;
804 memset(&ret.p->cnts, 0, sizeof(prof_cnt_t));
805 ret.p->gctx = gctx;
806 ret.p->tctx_uid = tdata->tctx_uid_next++;
807 ret.p->prepared = true;
808 ret.p->state = prof_tctx_state_initializing;
809 malloc_mutex_lock(tdata->lock);
810 error = ckh_insert(tsd, &tdata->bt2tctx, btkey, ret.v);
811 malloc_mutex_unlock(tdata->lock);
812 if (error) {
813 if (new_gctx)
814 prof_gctx_try_destroy(tsd, tdata, gctx, tdata);
815 idalloctm(tsd, ret.v, tcache, true, true);
816 return (NULL);
817 }
818 malloc_mutex_lock(gctx->lock);
819 ret.p->state = prof_tctx_state_nominal;
820 tctx_tree_insert(&gctx->tctxs, ret.p);
821 gctx->nlimbo--;
822 malloc_mutex_unlock(gctx->lock);
823 }
824
825 return (ret.p);
826 }
827
828 void
prof_sample_threshold_update(prof_tdata_t * tdata)829 prof_sample_threshold_update(prof_tdata_t *tdata)
830 {
831 /*
832 * The body of this function is compiled out unless heap profiling is
833 * enabled, so that it is possible to compile jemalloc with floating
834 * point support completely disabled. Avoiding floating point code is
835 * important on memory-constrained systems, but it also enables a
836 * workaround for versions of glibc that don't properly save/restore
837 * floating point registers during dynamic lazy symbol loading (which
838 * internally calls into whatever malloc implementation happens to be
839 * integrated into the application). Note that some compilers (e.g.
840 * gcc 4.8) may use floating point registers for fast memory moves, so
841 * jemalloc must be compiled with such optimizations disabled (e.g.
842 * -mno-sse) in order for the workaround to be complete.
843 */
844 #ifdef JEMALLOC_PROF
845 uint64_t r;
846 double u;
847
848 if (!config_prof)
849 return;
850
851 if (lg_prof_sample == 0) {
852 tdata->bytes_until_sample = 0;
853 return;
854 }
855
856 /*
857 * Compute sample interval as a geometrically distributed random
858 * variable with mean (2^lg_prof_sample).
859 *
860 * __ __
861 * | log(u) | 1
862 * tdata->bytes_until_sample = | -------- |, where p = ---------------
863 * | log(1-p) | lg_prof_sample
864 * 2
865 *
866 * For more information on the math, see:
867 *
868 * Non-Uniform Random Variate Generation
869 * Luc Devroye
870 * Springer-Verlag, New York, 1986
871 * pp 500
872 * (http://luc.devroye.org/rnbookindex.html)
873 */
874 r = prng_lg_range(&tdata->prng_state, 53);
875 u = (double)r * (1.0/9007199254740992.0L);
876 tdata->bytes_until_sample = (uint64_t)(log(u) /
877 log(1.0 - (1.0 / (double)((uint64_t)1U << lg_prof_sample))))
878 + (uint64_t)1U;
879 #endif
880 }
881
882 #ifdef JEMALLOC_JET
883 static prof_tdata_t *
prof_tdata_count_iter(prof_tdata_tree_t * tdatas,prof_tdata_t * tdata,void * arg)884 prof_tdata_count_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata, void *arg)
885 {
886 size_t *tdata_count = (size_t *)arg;
887
888 (*tdata_count)++;
889
890 return (NULL);
891 }
892
893 size_t
prof_tdata_count(void)894 prof_tdata_count(void)
895 {
896 size_t tdata_count = 0;
897
898 malloc_mutex_lock(&tdatas_mtx);
899 tdata_tree_iter(&tdatas, NULL, prof_tdata_count_iter,
900 (void *)&tdata_count);
901 malloc_mutex_unlock(&tdatas_mtx);
902
903 return (tdata_count);
904 }
905 #endif
906
907 #ifdef JEMALLOC_JET
908 size_t
prof_bt_count(void)909 prof_bt_count(void)
910 {
911 size_t bt_count;
912 tsd_t *tsd;
913 prof_tdata_t *tdata;
914
915 tsd = tsd_fetch();
916 tdata = prof_tdata_get(tsd, false);
917 if (tdata == NULL)
918 return (0);
919
920 malloc_mutex_lock(&bt2gctx_mtx);
921 bt_count = ckh_count(&bt2gctx);
922 malloc_mutex_unlock(&bt2gctx_mtx);
923
924 return (bt_count);
925 }
926 #endif
927
928 #ifdef JEMALLOC_JET
929 #undef prof_dump_open
930 #define prof_dump_open JEMALLOC_N(prof_dump_open_impl)
931 #endif
932 static int
prof_dump_open(bool propagate_err,const char * filename)933 prof_dump_open(bool propagate_err, const char *filename)
934 {
935 int fd;
936
937 fd = creat(filename, 0644);
938 if (fd == -1 && !propagate_err) {
939 malloc_printf("<jemalloc>: creat(\"%s\"), 0644) failed\n",
940 filename);
941 if (opt_abort)
942 abort();
943 }
944
945 return (fd);
946 }
947 #ifdef JEMALLOC_JET
948 #undef prof_dump_open
949 #define prof_dump_open JEMALLOC_N(prof_dump_open)
950 prof_dump_open_t *prof_dump_open = JEMALLOC_N(prof_dump_open_impl);
951 #endif
952
953 static bool
prof_dump_flush(bool propagate_err)954 prof_dump_flush(bool propagate_err)
955 {
956 bool ret = false;
957 ssize_t err;
958
959 cassert(config_prof);
960
961 err = write(prof_dump_fd, prof_dump_buf, prof_dump_buf_end);
962 if (err == -1) {
963 if (!propagate_err) {
964 malloc_write("<jemalloc>: write() failed during heap "
965 "profile flush\n");
966 if (opt_abort)
967 abort();
968 }
969 ret = true;
970 }
971 prof_dump_buf_end = 0;
972
973 return (ret);
974 }
975
976 static bool
prof_dump_close(bool propagate_err)977 prof_dump_close(bool propagate_err)
978 {
979 bool ret;
980
981 assert(prof_dump_fd != -1);
982 ret = prof_dump_flush(propagate_err);
983 close(prof_dump_fd);
984 prof_dump_fd = -1;
985
986 return (ret);
987 }
988
989 static bool
prof_dump_write(bool propagate_err,const char * s)990 prof_dump_write(bool propagate_err, const char *s)
991 {
992 size_t i, slen, n;
993
994 cassert(config_prof);
995
996 i = 0;
997 slen = strlen(s);
998 while (i < slen) {
999 /* Flush the buffer if it is full. */
1000 if (prof_dump_buf_end == PROF_DUMP_BUFSIZE)
1001 if (prof_dump_flush(propagate_err) && propagate_err)
1002 return (true);
1003
1004 if (prof_dump_buf_end + slen <= PROF_DUMP_BUFSIZE) {
1005 /* Finish writing. */
1006 n = slen - i;
1007 } else {
1008 /* Write as much of s as will fit. */
1009 n = PROF_DUMP_BUFSIZE - prof_dump_buf_end;
1010 }
1011 memcpy(&prof_dump_buf[prof_dump_buf_end], &s[i], n);
1012 prof_dump_buf_end += n;
1013 i += n;
1014 }
1015
1016 return (false);
1017 }
1018
1019 JEMALLOC_FORMAT_PRINTF(2, 3)
1020 static bool
prof_dump_printf(bool propagate_err,const char * format,...)1021 prof_dump_printf(bool propagate_err, const char *format, ...)
1022 {
1023 bool ret;
1024 va_list ap;
1025 char buf[PROF_PRINTF_BUFSIZE];
1026
1027 va_start(ap, format);
1028 malloc_vsnprintf(buf, sizeof(buf), format, ap);
1029 va_end(ap);
1030 ret = prof_dump_write(propagate_err, buf);
1031
1032 return (ret);
1033 }
1034
1035 /* tctx->tdata->lock is held. */
1036 static void
prof_tctx_merge_tdata(prof_tctx_t * tctx,prof_tdata_t * tdata)1037 prof_tctx_merge_tdata(prof_tctx_t *tctx, prof_tdata_t *tdata)
1038 {
1039
1040 malloc_mutex_lock(tctx->gctx->lock);
1041
1042 switch (tctx->state) {
1043 case prof_tctx_state_initializing:
1044 malloc_mutex_unlock(tctx->gctx->lock);
1045 return;
1046 case prof_tctx_state_nominal:
1047 tctx->state = prof_tctx_state_dumping;
1048 malloc_mutex_unlock(tctx->gctx->lock);
1049
1050 memcpy(&tctx->dump_cnts, &tctx->cnts, sizeof(prof_cnt_t));
1051
1052 tdata->cnt_summed.curobjs += tctx->dump_cnts.curobjs;
1053 tdata->cnt_summed.curbytes += tctx->dump_cnts.curbytes;
1054 if (opt_prof_accum) {
1055 tdata->cnt_summed.accumobjs +=
1056 tctx->dump_cnts.accumobjs;
1057 tdata->cnt_summed.accumbytes +=
1058 tctx->dump_cnts.accumbytes;
1059 }
1060 break;
1061 case prof_tctx_state_dumping:
1062 case prof_tctx_state_purgatory:
1063 not_reached();
1064 }
1065 }
1066
1067 /* gctx->lock is held. */
1068 static void
prof_tctx_merge_gctx(prof_tctx_t * tctx,prof_gctx_t * gctx)1069 prof_tctx_merge_gctx(prof_tctx_t *tctx, prof_gctx_t *gctx)
1070 {
1071
1072 gctx->cnt_summed.curobjs += tctx->dump_cnts.curobjs;
1073 gctx->cnt_summed.curbytes += tctx->dump_cnts.curbytes;
1074 if (opt_prof_accum) {
1075 gctx->cnt_summed.accumobjs += tctx->dump_cnts.accumobjs;
1076 gctx->cnt_summed.accumbytes += tctx->dump_cnts.accumbytes;
1077 }
1078 }
1079
1080 /* tctx->gctx is held. */
1081 static prof_tctx_t *
prof_tctx_merge_iter(prof_tctx_tree_t * tctxs,prof_tctx_t * tctx,void * arg)1082 prof_tctx_merge_iter(prof_tctx_tree_t *tctxs, prof_tctx_t *tctx, void *arg)
1083 {
1084
1085 switch (tctx->state) {
1086 case prof_tctx_state_nominal:
1087 /* New since dumping started; ignore. */
1088 break;
1089 case prof_tctx_state_dumping:
1090 case prof_tctx_state_purgatory:
1091 prof_tctx_merge_gctx(tctx, tctx->gctx);
1092 break;
1093 default:
1094 not_reached();
1095 }
1096
1097 return (NULL);
1098 }
1099
1100 /* gctx->lock is held. */
1101 static prof_tctx_t *
prof_tctx_dump_iter(prof_tctx_tree_t * tctxs,prof_tctx_t * tctx,void * arg)1102 prof_tctx_dump_iter(prof_tctx_tree_t *tctxs, prof_tctx_t *tctx, void *arg)
1103 {
1104 bool propagate_err = *(bool *)arg;
1105
1106 switch (tctx->state) {
1107 case prof_tctx_state_initializing:
1108 case prof_tctx_state_nominal:
1109 /* Not captured by this dump. */
1110 break;
1111 case prof_tctx_state_dumping:
1112 case prof_tctx_state_purgatory:
1113 if (prof_dump_printf(propagate_err,
1114 " t%"FMTu64": %"FMTu64": %"FMTu64" [%"FMTu64": "
1115 "%"FMTu64"]\n", tctx->thr_uid, tctx->dump_cnts.curobjs,
1116 tctx->dump_cnts.curbytes, tctx->dump_cnts.accumobjs,
1117 tctx->dump_cnts.accumbytes))
1118 return (tctx);
1119 break;
1120 default:
1121 not_reached();
1122 }
1123 return (NULL);
1124 }
1125
1126 /* tctx->gctx is held. */
1127 static prof_tctx_t *
prof_tctx_finish_iter(prof_tctx_tree_t * tctxs,prof_tctx_t * tctx,void * arg)1128 prof_tctx_finish_iter(prof_tctx_tree_t *tctxs, prof_tctx_t *tctx, void *arg)
1129 {
1130 prof_tctx_t *ret;
1131
1132 switch (tctx->state) {
1133 case prof_tctx_state_nominal:
1134 /* New since dumping started; ignore. */
1135 break;
1136 case prof_tctx_state_dumping:
1137 tctx->state = prof_tctx_state_nominal;
1138 break;
1139 case prof_tctx_state_purgatory:
1140 ret = tctx;
1141 goto label_return;
1142 default:
1143 not_reached();
1144 }
1145
1146 ret = NULL;
1147 label_return:
1148 return (ret);
1149 }
1150
1151 static void
prof_dump_gctx_prep(prof_gctx_t * gctx,prof_gctx_tree_t * gctxs)1152 prof_dump_gctx_prep(prof_gctx_t *gctx, prof_gctx_tree_t *gctxs)
1153 {
1154
1155 cassert(config_prof);
1156
1157 malloc_mutex_lock(gctx->lock);
1158
1159 /*
1160 * Increment nlimbo so that gctx won't go away before dump.
1161 * Additionally, link gctx into the dump list so that it is included in
1162 * prof_dump()'s second pass.
1163 */
1164 gctx->nlimbo++;
1165 gctx_tree_insert(gctxs, gctx);
1166
1167 memset(&gctx->cnt_summed, 0, sizeof(prof_cnt_t));
1168
1169 malloc_mutex_unlock(gctx->lock);
1170 }
1171
1172 static prof_gctx_t *
prof_gctx_merge_iter(prof_gctx_tree_t * gctxs,prof_gctx_t * gctx,void * arg)1173 prof_gctx_merge_iter(prof_gctx_tree_t *gctxs, prof_gctx_t *gctx, void *arg)
1174 {
1175 size_t *leak_ngctx = (size_t *)arg;
1176
1177 malloc_mutex_lock(gctx->lock);
1178 tctx_tree_iter(&gctx->tctxs, NULL, prof_tctx_merge_iter, NULL);
1179 if (gctx->cnt_summed.curobjs != 0)
1180 (*leak_ngctx)++;
1181 malloc_mutex_unlock(gctx->lock);
1182
1183 return (NULL);
1184 }
1185
1186 static void
prof_gctx_finish(tsd_t * tsd,prof_gctx_tree_t * gctxs)1187 prof_gctx_finish(tsd_t *tsd, prof_gctx_tree_t *gctxs)
1188 {
1189 prof_tdata_t *tdata = prof_tdata_get(tsd, false);
1190 prof_gctx_t *gctx;
1191
1192 /*
1193 * Standard tree iteration won't work here, because as soon as we
1194 * decrement gctx->nlimbo and unlock gctx, another thread can
1195 * concurrently destroy it, which will corrupt the tree. Therefore,
1196 * tear down the tree one node at a time during iteration.
1197 */
1198 while ((gctx = gctx_tree_first(gctxs)) != NULL) {
1199 gctx_tree_remove(gctxs, gctx);
1200 malloc_mutex_lock(gctx->lock);
1201 {
1202 prof_tctx_t *next;
1203
1204 next = NULL;
1205 do {
1206 prof_tctx_t *to_destroy =
1207 tctx_tree_iter(&gctx->tctxs, next,
1208 prof_tctx_finish_iter, NULL);
1209 if (to_destroy != NULL) {
1210 next = tctx_tree_next(&gctx->tctxs,
1211 to_destroy);
1212 tctx_tree_remove(&gctx->tctxs,
1213 to_destroy);
1214 idalloctm(tsd, to_destroy,
1215 tcache_get(tsd, false), true, true);
1216 } else
1217 next = NULL;
1218 } while (next != NULL);
1219 }
1220 gctx->nlimbo--;
1221 if (prof_gctx_should_destroy(gctx)) {
1222 gctx->nlimbo++;
1223 malloc_mutex_unlock(gctx->lock);
1224 prof_gctx_try_destroy(tsd, tdata, gctx, tdata);
1225 } else
1226 malloc_mutex_unlock(gctx->lock);
1227 }
1228 }
1229
1230 static prof_tdata_t *
prof_tdata_merge_iter(prof_tdata_tree_t * tdatas,prof_tdata_t * tdata,void * arg)1231 prof_tdata_merge_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata, void *arg)
1232 {
1233 prof_cnt_t *cnt_all = (prof_cnt_t *)arg;
1234
1235 malloc_mutex_lock(tdata->lock);
1236 if (!tdata->expired) {
1237 size_t tabind;
1238 union {
1239 prof_tctx_t *p;
1240 void *v;
1241 } tctx;
1242
1243 tdata->dumping = true;
1244 memset(&tdata->cnt_summed, 0, sizeof(prof_cnt_t));
1245 for (tabind = 0; !ckh_iter(&tdata->bt2tctx, &tabind, NULL,
1246 &tctx.v);)
1247 prof_tctx_merge_tdata(tctx.p, tdata);
1248
1249 cnt_all->curobjs += tdata->cnt_summed.curobjs;
1250 cnt_all->curbytes += tdata->cnt_summed.curbytes;
1251 if (opt_prof_accum) {
1252 cnt_all->accumobjs += tdata->cnt_summed.accumobjs;
1253 cnt_all->accumbytes += tdata->cnt_summed.accumbytes;
1254 }
1255 } else
1256 tdata->dumping = false;
1257 malloc_mutex_unlock(tdata->lock);
1258
1259 return (NULL);
1260 }
1261
1262 static prof_tdata_t *
prof_tdata_dump_iter(prof_tdata_tree_t * tdatas,prof_tdata_t * tdata,void * arg)1263 prof_tdata_dump_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata, void *arg)
1264 {
1265 bool propagate_err = *(bool *)arg;
1266
1267 if (!tdata->dumping)
1268 return (NULL);
1269
1270 if (prof_dump_printf(propagate_err,
1271 " t%"FMTu64": %"FMTu64": %"FMTu64" [%"FMTu64": %"FMTu64"]%s%s\n",
1272 tdata->thr_uid, tdata->cnt_summed.curobjs,
1273 tdata->cnt_summed.curbytes, tdata->cnt_summed.accumobjs,
1274 tdata->cnt_summed.accumbytes,
1275 (tdata->thread_name != NULL) ? " " : "",
1276 (tdata->thread_name != NULL) ? tdata->thread_name : ""))
1277 return (tdata);
1278 return (NULL);
1279 }
1280
1281 #ifdef JEMALLOC_JET
1282 #undef prof_dump_header
1283 #define prof_dump_header JEMALLOC_N(prof_dump_header_impl)
1284 #endif
1285 static bool
prof_dump_header(bool propagate_err,const prof_cnt_t * cnt_all)1286 prof_dump_header(bool propagate_err, const prof_cnt_t *cnt_all)
1287 {
1288 bool ret;
1289
1290 if (prof_dump_printf(propagate_err,
1291 "heap_v2/%"FMTu64"\n"
1292 " t*: %"FMTu64": %"FMTu64" [%"FMTu64": %"FMTu64"]\n",
1293 ((uint64_t)1U << lg_prof_sample), cnt_all->curobjs,
1294 cnt_all->curbytes, cnt_all->accumobjs, cnt_all->accumbytes))
1295 return (true);
1296
1297 malloc_mutex_lock(&tdatas_mtx);
1298 ret = (tdata_tree_iter(&tdatas, NULL, prof_tdata_dump_iter,
1299 (void *)&propagate_err) != NULL);
1300 malloc_mutex_unlock(&tdatas_mtx);
1301 return (ret);
1302 }
1303 #ifdef JEMALLOC_JET
1304 #undef prof_dump_header
1305 #define prof_dump_header JEMALLOC_N(prof_dump_header)
1306 prof_dump_header_t *prof_dump_header = JEMALLOC_N(prof_dump_header_impl);
1307 #endif
1308
1309 /* gctx->lock is held. */
1310 static bool
prof_dump_gctx(bool propagate_err,prof_gctx_t * gctx,const prof_bt_t * bt,prof_gctx_tree_t * gctxs)1311 prof_dump_gctx(bool propagate_err, prof_gctx_t *gctx, const prof_bt_t *bt,
1312 prof_gctx_tree_t *gctxs)
1313 {
1314 bool ret;
1315 unsigned i;
1316
1317 cassert(config_prof);
1318
1319 /* Avoid dumping such gctx's that have no useful data. */
1320 if ((!opt_prof_accum && gctx->cnt_summed.curobjs == 0) ||
1321 (opt_prof_accum && gctx->cnt_summed.accumobjs == 0)) {
1322 assert(gctx->cnt_summed.curobjs == 0);
1323 assert(gctx->cnt_summed.curbytes == 0);
1324 assert(gctx->cnt_summed.accumobjs == 0);
1325 assert(gctx->cnt_summed.accumbytes == 0);
1326 ret = false;
1327 goto label_return;
1328 }
1329
1330 if (prof_dump_printf(propagate_err, "@")) {
1331 ret = true;
1332 goto label_return;
1333 }
1334 for (i = 0; i < bt->len; i++) {
1335 if (prof_dump_printf(propagate_err, " %#"FMTxPTR,
1336 (uintptr_t)bt->vec[i])) {
1337 ret = true;
1338 goto label_return;
1339 }
1340 }
1341
1342 if (prof_dump_printf(propagate_err,
1343 "\n"
1344 " t*: %"FMTu64": %"FMTu64" [%"FMTu64": %"FMTu64"]\n",
1345 gctx->cnt_summed.curobjs, gctx->cnt_summed.curbytes,
1346 gctx->cnt_summed.accumobjs, gctx->cnt_summed.accumbytes)) {
1347 ret = true;
1348 goto label_return;
1349 }
1350
1351 if (tctx_tree_iter(&gctx->tctxs, NULL, prof_tctx_dump_iter,
1352 (void *)&propagate_err) != NULL) {
1353 ret = true;
1354 goto label_return;
1355 }
1356
1357 ret = false;
1358 label_return:
1359 return (ret);
1360 }
1361
1362 #ifndef _WIN32
1363 JEMALLOC_FORMAT_PRINTF(1, 2)
1364 static int
prof_open_maps(const char * format,...)1365 prof_open_maps(const char *format, ...)
1366 {
1367 int mfd;
1368 va_list ap;
1369 char filename[PATH_MAX + 1];
1370
1371 va_start(ap, format);
1372 malloc_vsnprintf(filename, sizeof(filename), format, ap);
1373 va_end(ap);
1374 mfd = open(filename, O_RDONLY);
1375
1376 return (mfd);
1377 }
1378 #endif
1379
1380 static int
prof_getpid(void)1381 prof_getpid(void)
1382 {
1383
1384 #ifdef _WIN32
1385 return (GetCurrentProcessId());
1386 #else
1387 return (getpid());
1388 #endif
1389 }
1390
1391 static bool
prof_dump_maps(bool propagate_err)1392 prof_dump_maps(bool propagate_err)
1393 {
1394 bool ret;
1395 int mfd;
1396
1397 cassert(config_prof);
1398 #ifdef __FreeBSD__
1399 mfd = prof_open_maps("/proc/curproc/map");
1400 #elif defined(_WIN32)
1401 mfd = -1; // Not implemented
1402 #else
1403 {
1404 int pid = prof_getpid();
1405
1406 mfd = prof_open_maps("/proc/%d/task/%d/maps", pid, pid);
1407 if (mfd == -1)
1408 mfd = prof_open_maps("/proc/%d/maps", pid);
1409 }
1410 #endif
1411 if (mfd != -1) {
1412 ssize_t nread;
1413
1414 if (prof_dump_write(propagate_err, "\nMAPPED_LIBRARIES:\n") &&
1415 propagate_err) {
1416 ret = true;
1417 goto label_return;
1418 }
1419 nread = 0;
1420 do {
1421 prof_dump_buf_end += nread;
1422 if (prof_dump_buf_end == PROF_DUMP_BUFSIZE) {
1423 /* Make space in prof_dump_buf before read(). */
1424 if (prof_dump_flush(propagate_err) &&
1425 propagate_err) {
1426 ret = true;
1427 goto label_return;
1428 }
1429 }
1430 nread = read(mfd, &prof_dump_buf[prof_dump_buf_end],
1431 PROF_DUMP_BUFSIZE - prof_dump_buf_end);
1432 } while (nread > 0);
1433 } else {
1434 ret = true;
1435 goto label_return;
1436 }
1437
1438 ret = false;
1439 label_return:
1440 if (mfd != -1)
1441 close(mfd);
1442 return (ret);
1443 }
1444
1445 static void
prof_leakcheck(const prof_cnt_t * cnt_all,size_t leak_ngctx,const char * filename)1446 prof_leakcheck(const prof_cnt_t *cnt_all, size_t leak_ngctx,
1447 const char *filename)
1448 {
1449
1450 if (cnt_all->curbytes != 0) {
1451 malloc_printf("<jemalloc>: Leak summary: %"FMTu64" byte%s, %"
1452 FMTu64" object%s, %zu context%s\n",
1453 cnt_all->curbytes, (cnt_all->curbytes != 1) ? "s" : "",
1454 cnt_all->curobjs, (cnt_all->curobjs != 1) ? "s" : "",
1455 leak_ngctx, (leak_ngctx != 1) ? "s" : "");
1456 malloc_printf(
1457 "<jemalloc>: Run jeprof on \"%s\" for leak detail\n",
1458 filename);
1459 }
1460 }
1461
1462 static prof_gctx_t *
prof_gctx_dump_iter(prof_gctx_tree_t * gctxs,prof_gctx_t * gctx,void * arg)1463 prof_gctx_dump_iter(prof_gctx_tree_t *gctxs, prof_gctx_t *gctx, void *arg)
1464 {
1465 prof_gctx_t *ret;
1466 bool propagate_err = *(bool *)arg;
1467
1468 malloc_mutex_lock(gctx->lock);
1469
1470 if (prof_dump_gctx(propagate_err, gctx, &gctx->bt, gctxs)) {
1471 ret = gctx;
1472 goto label_return;
1473 }
1474
1475 ret = NULL;
1476 label_return:
1477 malloc_mutex_unlock(gctx->lock);
1478 return (ret);
1479 }
1480
1481 static bool
prof_dump(tsd_t * tsd,bool propagate_err,const char * filename,bool leakcheck)1482 prof_dump(tsd_t *tsd, bool propagate_err, const char *filename, bool leakcheck)
1483 {
1484 prof_tdata_t *tdata;
1485 prof_cnt_t cnt_all;
1486 size_t tabind;
1487 union {
1488 prof_gctx_t *p;
1489 void *v;
1490 } gctx;
1491 size_t leak_ngctx;
1492 prof_gctx_tree_t gctxs;
1493
1494 cassert(config_prof);
1495
1496 tdata = prof_tdata_get(tsd, true);
1497 if (tdata == NULL)
1498 return (true);
1499
1500 malloc_mutex_lock(&prof_dump_mtx);
1501 prof_enter(tsd, tdata);
1502
1503 /*
1504 * Put gctx's in limbo and clear their counters in preparation for
1505 * summing.
1506 */
1507 gctx_tree_new(&gctxs);
1508 for (tabind = 0; !ckh_iter(&bt2gctx, &tabind, NULL, &gctx.v);)
1509 prof_dump_gctx_prep(gctx.p, &gctxs);
1510
1511 /*
1512 * Iterate over tdatas, and for the non-expired ones snapshot their tctx
1513 * stats and merge them into the associated gctx's.
1514 */
1515 memset(&cnt_all, 0, sizeof(prof_cnt_t));
1516 malloc_mutex_lock(&tdatas_mtx);
1517 tdata_tree_iter(&tdatas, NULL, prof_tdata_merge_iter, (void *)&cnt_all);
1518 malloc_mutex_unlock(&tdatas_mtx);
1519
1520 /* Merge tctx stats into gctx's. */
1521 leak_ngctx = 0;
1522 gctx_tree_iter(&gctxs, NULL, prof_gctx_merge_iter, (void *)&leak_ngctx);
1523
1524 prof_leave(tsd, tdata);
1525
1526 /* Create dump file. */
1527 if ((prof_dump_fd = prof_dump_open(propagate_err, filename)) == -1)
1528 goto label_open_close_error;
1529
1530 /* Dump profile header. */
1531 if (prof_dump_header(propagate_err, &cnt_all))
1532 goto label_write_error;
1533
1534 /* Dump per gctx profile stats. */
1535 if (gctx_tree_iter(&gctxs, NULL, prof_gctx_dump_iter,
1536 (void *)&propagate_err) != NULL)
1537 goto label_write_error;
1538
1539 /* Dump /proc/<pid>/maps if possible. */
1540 if (prof_dump_maps(propagate_err))
1541 goto label_write_error;
1542
1543 if (prof_dump_close(propagate_err))
1544 goto label_open_close_error;
1545
1546 prof_gctx_finish(tsd, &gctxs);
1547 malloc_mutex_unlock(&prof_dump_mtx);
1548
1549 if (leakcheck)
1550 prof_leakcheck(&cnt_all, leak_ngctx, filename);
1551
1552 return (false);
1553 label_write_error:
1554 prof_dump_close(propagate_err);
1555 label_open_close_error:
1556 prof_gctx_finish(tsd, &gctxs);
1557 malloc_mutex_unlock(&prof_dump_mtx);
1558 return (true);
1559 }
1560
1561 #define DUMP_FILENAME_BUFSIZE (PATH_MAX + 1)
1562 #define VSEQ_INVALID UINT64_C(0xffffffffffffffff)
1563 static void
prof_dump_filename(char * filename,char v,uint64_t vseq)1564 prof_dump_filename(char *filename, char v, uint64_t vseq)
1565 {
1566
1567 cassert(config_prof);
1568
1569 if (vseq != VSEQ_INVALID) {
1570 /* "<prefix>.<pid>.<seq>.v<vseq>.heap" */
1571 malloc_snprintf(filename, DUMP_FILENAME_BUFSIZE,
1572 "%s.%d.%"FMTu64".%c%"FMTu64".heap",
1573 opt_prof_prefix, prof_getpid(), prof_dump_seq, v, vseq);
1574 } else {
1575 /* "<prefix>.<pid>.<seq>.<v>.heap" */
1576 malloc_snprintf(filename, DUMP_FILENAME_BUFSIZE,
1577 "%s.%d.%"FMTu64".%c.heap",
1578 opt_prof_prefix, prof_getpid(), prof_dump_seq, v);
1579 }
1580 prof_dump_seq++;
1581 }
1582
1583 static void
prof_fdump(void)1584 prof_fdump(void)
1585 {
1586 tsd_t *tsd;
1587 char filename[DUMP_FILENAME_BUFSIZE];
1588
1589 cassert(config_prof);
1590 assert(opt_prof_final);
1591 assert(opt_prof_prefix[0] != '\0');
1592
1593 if (!prof_booted)
1594 return;
1595 tsd = tsd_fetch();
1596
1597 malloc_mutex_lock(&prof_dump_seq_mtx);
1598 prof_dump_filename(filename, 'f', VSEQ_INVALID);
1599 malloc_mutex_unlock(&prof_dump_seq_mtx);
1600 prof_dump(tsd, false, filename, opt_prof_leak);
1601 }
1602
1603 void
prof_idump(void)1604 prof_idump(void)
1605 {
1606 tsd_t *tsd;
1607 prof_tdata_t *tdata;
1608
1609 cassert(config_prof);
1610
1611 if (!prof_booted)
1612 return;
1613 tsd = tsd_fetch();
1614 tdata = prof_tdata_get(tsd, false);
1615 if (tdata == NULL)
1616 return;
1617 if (tdata->enq) {
1618 tdata->enq_idump = true;
1619 return;
1620 }
1621
1622 if (opt_prof_prefix[0] != '\0') {
1623 char filename[PATH_MAX + 1];
1624 malloc_mutex_lock(&prof_dump_seq_mtx);
1625 prof_dump_filename(filename, 'i', prof_dump_iseq);
1626 prof_dump_iseq++;
1627 malloc_mutex_unlock(&prof_dump_seq_mtx);
1628 prof_dump(tsd, false, filename, false);
1629 }
1630 }
1631
1632 bool
prof_mdump(const char * filename)1633 prof_mdump(const char *filename)
1634 {
1635 tsd_t *tsd;
1636 char filename_buf[DUMP_FILENAME_BUFSIZE];
1637
1638 cassert(config_prof);
1639
1640 if (!opt_prof || !prof_booted)
1641 return (true);
1642 tsd = tsd_fetch();
1643
1644 if (filename == NULL) {
1645 /* No filename specified, so automatically generate one. */
1646 if (opt_prof_prefix[0] == '\0')
1647 return (true);
1648 malloc_mutex_lock(&prof_dump_seq_mtx);
1649 prof_dump_filename(filename_buf, 'm', prof_dump_mseq);
1650 prof_dump_mseq++;
1651 malloc_mutex_unlock(&prof_dump_seq_mtx);
1652 filename = filename_buf;
1653 }
1654 return (prof_dump(tsd, true, filename, false));
1655 }
1656
1657 void
prof_gdump(void)1658 prof_gdump(void)
1659 {
1660 tsd_t *tsd;
1661 prof_tdata_t *tdata;
1662
1663 cassert(config_prof);
1664
1665 if (!prof_booted)
1666 return;
1667 tsd = tsd_fetch();
1668 tdata = prof_tdata_get(tsd, false);
1669 if (tdata == NULL)
1670 return;
1671 if (tdata->enq) {
1672 tdata->enq_gdump = true;
1673 return;
1674 }
1675
1676 if (opt_prof_prefix[0] != '\0') {
1677 char filename[DUMP_FILENAME_BUFSIZE];
1678 malloc_mutex_lock(&prof_dump_seq_mtx);
1679 prof_dump_filename(filename, 'u', prof_dump_useq);
1680 prof_dump_useq++;
1681 malloc_mutex_unlock(&prof_dump_seq_mtx);
1682 prof_dump(tsd, false, filename, false);
1683 }
1684 }
1685
1686 static void
prof_bt_hash(const void * key,size_t r_hash[2])1687 prof_bt_hash(const void *key, size_t r_hash[2])
1688 {
1689 prof_bt_t *bt = (prof_bt_t *)key;
1690
1691 cassert(config_prof);
1692
1693 hash(bt->vec, bt->len * sizeof(void *), 0x94122f33U, r_hash);
1694 }
1695
1696 static bool
prof_bt_keycomp(const void * k1,const void * k2)1697 prof_bt_keycomp(const void *k1, const void *k2)
1698 {
1699 const prof_bt_t *bt1 = (prof_bt_t *)k1;
1700 const prof_bt_t *bt2 = (prof_bt_t *)k2;
1701
1702 cassert(config_prof);
1703
1704 if (bt1->len != bt2->len)
1705 return (false);
1706 return (memcmp(bt1->vec, bt2->vec, bt1->len * sizeof(void *)) == 0);
1707 }
1708
1709 JEMALLOC_INLINE_C uint64_t
prof_thr_uid_alloc(void)1710 prof_thr_uid_alloc(void)
1711 {
1712 uint64_t thr_uid;
1713
1714 malloc_mutex_lock(&next_thr_uid_mtx);
1715 thr_uid = next_thr_uid;
1716 next_thr_uid++;
1717 malloc_mutex_unlock(&next_thr_uid_mtx);
1718
1719 return (thr_uid);
1720 }
1721
1722 static prof_tdata_t *
prof_tdata_init_impl(tsd_t * tsd,uint64_t thr_uid,uint64_t thr_discrim,char * thread_name,bool active)1723 prof_tdata_init_impl(tsd_t *tsd, uint64_t thr_uid, uint64_t thr_discrim,
1724 char *thread_name, bool active)
1725 {
1726 prof_tdata_t *tdata;
1727 tcache_t *tcache;
1728
1729 cassert(config_prof);
1730
1731 /* Initialize an empty cache for this thread. */
1732 tcache = tcache_get(tsd, true);
1733 tdata = (prof_tdata_t *)iallocztm(tsd, sizeof(prof_tdata_t),
1734 size2index(sizeof(prof_tdata_t)), false, tcache, true, NULL, true);
1735 if (tdata == NULL)
1736 return (NULL);
1737
1738 tdata->lock = prof_tdata_mutex_choose(thr_uid);
1739 tdata->thr_uid = thr_uid;
1740 tdata->thr_discrim = thr_discrim;
1741 tdata->thread_name = thread_name;
1742 tdata->attached = true;
1743 tdata->expired = false;
1744 tdata->tctx_uid_next = 0;
1745
1746 if (ckh_new(tsd, &tdata->bt2tctx, PROF_CKH_MINITEMS,
1747 prof_bt_hash, prof_bt_keycomp)) {
1748 idalloctm(tsd, tdata, tcache, true, true);
1749 return (NULL);
1750 }
1751
1752 tdata->prng_state = (uint64_t)(uintptr_t)tdata;
1753 prof_sample_threshold_update(tdata);
1754
1755 tdata->enq = false;
1756 tdata->enq_idump = false;
1757 tdata->enq_gdump = false;
1758
1759 tdata->dumping = false;
1760 tdata->active = active;
1761
1762 malloc_mutex_lock(&tdatas_mtx);
1763 tdata_tree_insert(&tdatas, tdata);
1764 malloc_mutex_unlock(&tdatas_mtx);
1765
1766 return (tdata);
1767 }
1768
1769 prof_tdata_t *
prof_tdata_init(tsd_t * tsd)1770 prof_tdata_init(tsd_t *tsd)
1771 {
1772
1773 return (prof_tdata_init_impl(tsd, prof_thr_uid_alloc(), 0, NULL,
1774 prof_thread_active_init_get()));
1775 }
1776
1777 /* tdata->lock must be held. */
1778 static bool
prof_tdata_should_destroy(prof_tdata_t * tdata,bool even_if_attached)1779 prof_tdata_should_destroy(prof_tdata_t *tdata, bool even_if_attached)
1780 {
1781
1782 if (tdata->attached && !even_if_attached)
1783 return (false);
1784 if (ckh_count(&tdata->bt2tctx) != 0)
1785 return (false);
1786 return (true);
1787 }
1788
1789 /* tdatas_mtx must be held. */
1790 static void
prof_tdata_destroy_locked(tsd_t * tsd,prof_tdata_t * tdata,bool even_if_attached)1791 prof_tdata_destroy_locked(tsd_t *tsd, prof_tdata_t *tdata,
1792 bool even_if_attached)
1793 {
1794 tcache_t *tcache;
1795
1796 assert(prof_tdata_should_destroy(tdata, even_if_attached));
1797 assert(tsd_prof_tdata_get(tsd) != tdata);
1798
1799 tdata_tree_remove(&tdatas, tdata);
1800
1801 tcache = tcache_get(tsd, false);
1802 if (tdata->thread_name != NULL)
1803 idalloctm(tsd, tdata->thread_name, tcache, true, true);
1804 ckh_delete(tsd, &tdata->bt2tctx);
1805 idalloctm(tsd, tdata, tcache, true, true);
1806 }
1807
1808 static void
prof_tdata_destroy(tsd_t * tsd,prof_tdata_t * tdata,bool even_if_attached)1809 prof_tdata_destroy(tsd_t *tsd, prof_tdata_t *tdata, bool even_if_attached)
1810 {
1811
1812 malloc_mutex_lock(&tdatas_mtx);
1813 prof_tdata_destroy_locked(tsd, tdata, even_if_attached);
1814 malloc_mutex_unlock(&tdatas_mtx);
1815 }
1816
1817 static void
prof_tdata_detach(tsd_t * tsd,prof_tdata_t * tdata)1818 prof_tdata_detach(tsd_t *tsd, prof_tdata_t *tdata)
1819 {
1820 bool destroy_tdata;
1821
1822 malloc_mutex_lock(tdata->lock);
1823 if (tdata->attached) {
1824 destroy_tdata = prof_tdata_should_destroy(tdata, true);
1825 /*
1826 * Only detach if !destroy_tdata, because detaching would allow
1827 * another thread to win the race to destroy tdata.
1828 */
1829 if (!destroy_tdata)
1830 tdata->attached = false;
1831 tsd_prof_tdata_set(tsd, NULL);
1832 } else
1833 destroy_tdata = false;
1834 malloc_mutex_unlock(tdata->lock);
1835 if (destroy_tdata)
1836 prof_tdata_destroy(tsd, tdata, true);
1837 }
1838
1839 prof_tdata_t *
prof_tdata_reinit(tsd_t * tsd,prof_tdata_t * tdata)1840 prof_tdata_reinit(tsd_t *tsd, prof_tdata_t *tdata)
1841 {
1842 uint64_t thr_uid = tdata->thr_uid;
1843 uint64_t thr_discrim = tdata->thr_discrim + 1;
1844 char *thread_name = (tdata->thread_name != NULL) ?
1845 prof_thread_name_alloc(tsd, tdata->thread_name) : NULL;
1846 bool active = tdata->active;
1847
1848 prof_tdata_detach(tsd, tdata);
1849 return (prof_tdata_init_impl(tsd, thr_uid, thr_discrim, thread_name,
1850 active));
1851 }
1852
1853 static bool
prof_tdata_expire(prof_tdata_t * tdata)1854 prof_tdata_expire(prof_tdata_t *tdata)
1855 {
1856 bool destroy_tdata;
1857
1858 malloc_mutex_lock(tdata->lock);
1859 if (!tdata->expired) {
1860 tdata->expired = true;
1861 destroy_tdata = tdata->attached ? false :
1862 prof_tdata_should_destroy(tdata, false);
1863 } else
1864 destroy_tdata = false;
1865 malloc_mutex_unlock(tdata->lock);
1866
1867 return (destroy_tdata);
1868 }
1869
1870 static prof_tdata_t *
prof_tdata_reset_iter(prof_tdata_tree_t * tdatas,prof_tdata_t * tdata,void * arg)1871 prof_tdata_reset_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata, void *arg)
1872 {
1873
1874 return (prof_tdata_expire(tdata) ? tdata : NULL);
1875 }
1876
1877 void
prof_reset(tsd_t * tsd,size_t lg_sample)1878 prof_reset(tsd_t *tsd, size_t lg_sample)
1879 {
1880 prof_tdata_t *next;
1881
1882 assert(lg_sample < (sizeof(uint64_t) << 3));
1883
1884 malloc_mutex_lock(&prof_dump_mtx);
1885 malloc_mutex_lock(&tdatas_mtx);
1886
1887 lg_prof_sample = lg_sample;
1888
1889 next = NULL;
1890 do {
1891 prof_tdata_t *to_destroy = tdata_tree_iter(&tdatas, next,
1892 prof_tdata_reset_iter, NULL);
1893 if (to_destroy != NULL) {
1894 next = tdata_tree_next(&tdatas, to_destroy);
1895 prof_tdata_destroy_locked(tsd, to_destroy, false);
1896 } else
1897 next = NULL;
1898 } while (next != NULL);
1899
1900 malloc_mutex_unlock(&tdatas_mtx);
1901 malloc_mutex_unlock(&prof_dump_mtx);
1902 }
1903
1904 void
prof_tdata_cleanup(tsd_t * tsd)1905 prof_tdata_cleanup(tsd_t *tsd)
1906 {
1907 prof_tdata_t *tdata;
1908
1909 if (!config_prof)
1910 return;
1911
1912 tdata = tsd_prof_tdata_get(tsd);
1913 if (tdata != NULL)
1914 prof_tdata_detach(tsd, tdata);
1915 }
1916
1917 bool
prof_active_get(void)1918 prof_active_get(void)
1919 {
1920 bool prof_active_current;
1921
1922 malloc_mutex_lock(&prof_active_mtx);
1923 prof_active_current = prof_active;
1924 malloc_mutex_unlock(&prof_active_mtx);
1925 return (prof_active_current);
1926 }
1927
1928 bool
prof_active_set(bool active)1929 prof_active_set(bool active)
1930 {
1931 bool prof_active_old;
1932
1933 malloc_mutex_lock(&prof_active_mtx);
1934 prof_active_old = prof_active;
1935 prof_active = active;
1936 malloc_mutex_unlock(&prof_active_mtx);
1937 return (prof_active_old);
1938 }
1939
1940 const char *
prof_thread_name_get(void)1941 prof_thread_name_get(void)
1942 {
1943 tsd_t *tsd;
1944 prof_tdata_t *tdata;
1945
1946 tsd = tsd_fetch();
1947 tdata = prof_tdata_get(tsd, true);
1948 if (tdata == NULL)
1949 return ("");
1950 return (tdata->thread_name != NULL ? tdata->thread_name : "");
1951 }
1952
1953 static char *
prof_thread_name_alloc(tsd_t * tsd,const char * thread_name)1954 prof_thread_name_alloc(tsd_t *tsd, const char *thread_name)
1955 {
1956 char *ret;
1957 size_t size;
1958
1959 if (thread_name == NULL)
1960 return (NULL);
1961
1962 size = strlen(thread_name) + 1;
1963 if (size == 1)
1964 return ("");
1965
1966 ret = iallocztm(tsd, size, size2index(size), false, tcache_get(tsd,
1967 true), true, NULL, true);
1968 if (ret == NULL)
1969 return (NULL);
1970 memcpy(ret, thread_name, size);
1971 return (ret);
1972 }
1973
1974 int
prof_thread_name_set(tsd_t * tsd,const char * thread_name)1975 prof_thread_name_set(tsd_t *tsd, const char *thread_name)
1976 {
1977 prof_tdata_t *tdata;
1978 unsigned i;
1979 char *s;
1980
1981 tdata = prof_tdata_get(tsd, true);
1982 if (tdata == NULL)
1983 return (EAGAIN);
1984
1985 /* Validate input. */
1986 if (thread_name == NULL)
1987 return (EFAULT);
1988 for (i = 0; thread_name[i] != '\0'; i++) {
1989 char c = thread_name[i];
1990 if (!isgraph(c) && !isblank(c))
1991 return (EFAULT);
1992 }
1993
1994 s = prof_thread_name_alloc(tsd, thread_name);
1995 if (s == NULL)
1996 return (EAGAIN);
1997
1998 if (tdata->thread_name != NULL) {
1999 idalloctm(tsd, tdata->thread_name, tcache_get(tsd, false),
2000 true, true);
2001 tdata->thread_name = NULL;
2002 }
2003 if (strlen(s) > 0)
2004 tdata->thread_name = s;
2005 return (0);
2006 }
2007
2008 bool
prof_thread_active_get(void)2009 prof_thread_active_get(void)
2010 {
2011 tsd_t *tsd;
2012 prof_tdata_t *tdata;
2013
2014 tsd = tsd_fetch();
2015 tdata = prof_tdata_get(tsd, true);
2016 if (tdata == NULL)
2017 return (false);
2018 return (tdata->active);
2019 }
2020
2021 bool
prof_thread_active_set(bool active)2022 prof_thread_active_set(bool active)
2023 {
2024 tsd_t *tsd;
2025 prof_tdata_t *tdata;
2026
2027 tsd = tsd_fetch();
2028 tdata = prof_tdata_get(tsd, true);
2029 if (tdata == NULL)
2030 return (true);
2031 tdata->active = active;
2032 return (false);
2033 }
2034
2035 bool
prof_thread_active_init_get(void)2036 prof_thread_active_init_get(void)
2037 {
2038 bool active_init;
2039
2040 malloc_mutex_lock(&prof_thread_active_init_mtx);
2041 active_init = prof_thread_active_init;
2042 malloc_mutex_unlock(&prof_thread_active_init_mtx);
2043 return (active_init);
2044 }
2045
2046 bool
prof_thread_active_init_set(bool active_init)2047 prof_thread_active_init_set(bool active_init)
2048 {
2049 bool active_init_old;
2050
2051 malloc_mutex_lock(&prof_thread_active_init_mtx);
2052 active_init_old = prof_thread_active_init;
2053 prof_thread_active_init = active_init;
2054 malloc_mutex_unlock(&prof_thread_active_init_mtx);
2055 return (active_init_old);
2056 }
2057
2058 bool
prof_gdump_get(void)2059 prof_gdump_get(void)
2060 {
2061 bool prof_gdump_current;
2062
2063 malloc_mutex_lock(&prof_gdump_mtx);
2064 prof_gdump_current = prof_gdump_val;
2065 malloc_mutex_unlock(&prof_gdump_mtx);
2066 return (prof_gdump_current);
2067 }
2068
2069 bool
prof_gdump_set(bool gdump)2070 prof_gdump_set(bool gdump)
2071 {
2072 bool prof_gdump_old;
2073
2074 malloc_mutex_lock(&prof_gdump_mtx);
2075 prof_gdump_old = prof_gdump_val;
2076 prof_gdump_val = gdump;
2077 malloc_mutex_unlock(&prof_gdump_mtx);
2078 return (prof_gdump_old);
2079 }
2080
2081 void
prof_boot0(void)2082 prof_boot0(void)
2083 {
2084
2085 cassert(config_prof);
2086
2087 memcpy(opt_prof_prefix, PROF_PREFIX_DEFAULT,
2088 sizeof(PROF_PREFIX_DEFAULT));
2089 }
2090
2091 void
prof_boot1(void)2092 prof_boot1(void)
2093 {
2094
2095 cassert(config_prof);
2096
2097 /*
2098 * opt_prof must be in its final state before any arenas are
2099 * initialized, so this function must be executed early.
2100 */
2101
2102 if (opt_prof_leak && !opt_prof) {
2103 /*
2104 * Enable opt_prof, but in such a way that profiles are never
2105 * automatically dumped.
2106 */
2107 opt_prof = true;
2108 opt_prof_gdump = false;
2109 } else if (opt_prof) {
2110 if (opt_lg_prof_interval >= 0) {
2111 prof_interval = (((uint64_t)1U) <<
2112 opt_lg_prof_interval);
2113 }
2114 }
2115 }
2116
2117 bool
prof_boot2(void)2118 prof_boot2(void)
2119 {
2120
2121 cassert(config_prof);
2122
2123 if (opt_prof) {
2124 tsd_t *tsd;
2125 unsigned i;
2126
2127 lg_prof_sample = opt_lg_prof_sample;
2128
2129 prof_active = opt_prof_active;
2130 if (malloc_mutex_init(&prof_active_mtx))
2131 return (true);
2132
2133 prof_gdump_val = opt_prof_gdump;
2134 if (malloc_mutex_init(&prof_gdump_mtx))
2135 return (true);
2136
2137 prof_thread_active_init = opt_prof_thread_active_init;
2138 if (malloc_mutex_init(&prof_thread_active_init_mtx))
2139 return (true);
2140
2141 tsd = tsd_fetch();
2142 if (ckh_new(tsd, &bt2gctx, PROF_CKH_MINITEMS, prof_bt_hash,
2143 prof_bt_keycomp))
2144 return (true);
2145 if (malloc_mutex_init(&bt2gctx_mtx))
2146 return (true);
2147
2148 tdata_tree_new(&tdatas);
2149 if (malloc_mutex_init(&tdatas_mtx))
2150 return (true);
2151
2152 next_thr_uid = 0;
2153 if (malloc_mutex_init(&next_thr_uid_mtx))
2154 return (true);
2155
2156 if (malloc_mutex_init(&prof_dump_seq_mtx))
2157 return (true);
2158 if (malloc_mutex_init(&prof_dump_mtx))
2159 return (true);
2160
2161 if (opt_prof_final && opt_prof_prefix[0] != '\0' &&
2162 atexit(prof_fdump) != 0) {
2163 malloc_write("<jemalloc>: Error in atexit()\n");
2164 if (opt_abort)
2165 abort();
2166 }
2167
2168 gctx_locks = (malloc_mutex_t *)base_alloc(PROF_NCTX_LOCKS *
2169 sizeof(malloc_mutex_t));
2170 if (gctx_locks == NULL)
2171 return (true);
2172 for (i = 0; i < PROF_NCTX_LOCKS; i++) {
2173 if (malloc_mutex_init(&gctx_locks[i]))
2174 return (true);
2175 }
2176
2177 tdata_locks = (malloc_mutex_t *)base_alloc(PROF_NTDATA_LOCKS *
2178 sizeof(malloc_mutex_t));
2179 if (tdata_locks == NULL)
2180 return (true);
2181 for (i = 0; i < PROF_NTDATA_LOCKS; i++) {
2182 if (malloc_mutex_init(&tdata_locks[i]))
2183 return (true);
2184 }
2185 }
2186
2187 #ifdef JEMALLOC_PROF_LIBGCC
2188 /*
2189 * Cause the backtracing machinery to allocate its internal state
2190 * before enabling profiling.
2191 */
2192 _Unwind_Backtrace(prof_unwind_init_callback, NULL);
2193 #endif
2194
2195 prof_booted = true;
2196
2197 return (false);
2198 }
2199
2200 void
prof_prefork(void)2201 prof_prefork(void)
2202 {
2203
2204 if (opt_prof) {
2205 unsigned i;
2206
2207 malloc_mutex_prefork(&tdatas_mtx);
2208 malloc_mutex_prefork(&bt2gctx_mtx);
2209 malloc_mutex_prefork(&next_thr_uid_mtx);
2210 malloc_mutex_prefork(&prof_dump_seq_mtx);
2211 for (i = 0; i < PROF_NCTX_LOCKS; i++)
2212 malloc_mutex_prefork(&gctx_locks[i]);
2213 for (i = 0; i < PROF_NTDATA_LOCKS; i++)
2214 malloc_mutex_prefork(&tdata_locks[i]);
2215 }
2216 }
2217
2218 void
prof_postfork_parent(void)2219 prof_postfork_parent(void)
2220 {
2221
2222 if (opt_prof) {
2223 unsigned i;
2224
2225 for (i = 0; i < PROF_NTDATA_LOCKS; i++)
2226 malloc_mutex_postfork_parent(&tdata_locks[i]);
2227 for (i = 0; i < PROF_NCTX_LOCKS; i++)
2228 malloc_mutex_postfork_parent(&gctx_locks[i]);
2229 malloc_mutex_postfork_parent(&prof_dump_seq_mtx);
2230 malloc_mutex_postfork_parent(&next_thr_uid_mtx);
2231 malloc_mutex_postfork_parent(&bt2gctx_mtx);
2232 malloc_mutex_postfork_parent(&tdatas_mtx);
2233 }
2234 }
2235
2236 void
prof_postfork_child(void)2237 prof_postfork_child(void)
2238 {
2239
2240 if (opt_prof) {
2241 unsigned i;
2242
2243 for (i = 0; i < PROF_NTDATA_LOCKS; i++)
2244 malloc_mutex_postfork_child(&tdata_locks[i]);
2245 for (i = 0; i < PROF_NCTX_LOCKS; i++)
2246 malloc_mutex_postfork_child(&gctx_locks[i]);
2247 malloc_mutex_postfork_child(&prof_dump_seq_mtx);
2248 malloc_mutex_postfork_child(&next_thr_uid_mtx);
2249 malloc_mutex_postfork_child(&bt2gctx_mtx);
2250 malloc_mutex_postfork_child(&tdatas_mtx);
2251 }
2252 }
2253
2254 /******************************************************************************/
2255